ROCm-Developer-Tools
[Feature]: Renaming signal op
Suggestion Description
Description
Since overal strategy of ROCM library in past, developers need to have a closer look at NVIDIA nvshem implementation (shareing the software development fundations).
Currently signal op such as nvshmemx_signal_op is missed from library.
We have to replace warp-level signal op with thread-level signal op.
For example to implement rocshmemx_signal_op which does not exist in rocshmem we have written codes such as
...
rocshmem_int_p(dest, value, target_pe);
...
Solution
Implement rocshmem_int_p based signal op to immedaitely catch up the nvshmem signal op interface.
Full signature :
device inline void rocshmem_signal_op(uint64_t *sig_addr, uint64_t signal, int sig_op, int pe)
Operating System
No response
GPU
No response
ROCm Component
No response
Hello,
Your proposed rocshmem_int_p workaround is incorrect and will produce randomly wrong results (from atomicity race condition).
A more correct workaround would use rocshmem_putmem_signal with 0-byte nelems. Note that _wg and _wave variants (equivalent to _block and _warp) are also available.
@abouteiller thanks ! You are right ! I will update with function name!